Title: SPADE: Sketch-guided Path Planning Augmented with Diffusion Experts

Abstract:

Effective path planning is a critical component for Autonomous Mobile Robots (AMRs). While integrating human preferences into planning algorithms is standard practice, traditional approaches often depend on intricate reward engineering or require substantial hardware resources. Recently, state-of-the-art frameworks have adopted imitation learning to develop behavior-specific path planning models trained on expert demonstrations. Despite this progress, such methods suffer from two primary drawbacks: poor generalization capabilities in unfamiliar environments and low robustness during the collection of demonstration data.

To overcome these hurdles, this study presents an improved framework centered on two significant contributions. First, it features a revamped annotation tool constructed on the ROS 2 platform. Second, it introduces a novel training methodology that incorporates diffusion-based augmentation into standard behavioral cloning models. The paper provides a dataset of expert demonstrations and validates the proposed solution’s robustness through comprehensive ablation studies.

The enhanced method demonstrates superior performance compared to existing state-of-the-art techniques, achieving a 39.1% reduction in Absolute Pose Error (APE) and a 33.5% decrease in Fréchet Inception Distance (FID). Notably, this improvement is realized with 93.8% fewer trainable parameters. Furthermore, the approach successfully achieves generalization capabilities comparable to diffusion models while maintaining the real-time, on-edge processing characteristics inherent to current state-of-the-art models.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC